The SpatioTemporalTask currently can handle generative models that do multiple predictions per sample. However most of our models are deterministic and hence this functionality makes the validation & test loop code less readable and more error prone.
Goal:
Refactor Stochastic predictions as a special case, that by default is not used, but can be if necessary.
Refactor validation & test loops to the specific case of deterministic prediction with the possibility to, if necessary, switch to the stochastic case
Culprits:
Currently the more general, stochastic, case is implemented, so it might be difficult to fall back to a more specific case?
The
SpatioTemporalTaskcurrently can handle generative models that do multiple predictions per sample. However most of our models are deterministic and hence this functionality makes the validation & test loop code less readable and more error prone.Goal:
Culprits: